Recommendation systems have been machine learning for over a decade. The AI era added better embeddings, sequence models, and conversational recommendation. Production systems combine multiple approaches. This post is the patterns that dominate production recommendation in 2026 and what the LLM era specifically changed.
Content-based filtering
Recommend items similar to what user interacted with. 'If you liked X, you'll like Y because Y has similar features.'
Works for new items. Features known even without interaction data.
Cold-start friendly. Works for items with no interaction history yet.
Limitations. Restricted to items with known features. Doesn't capture taste beyond explicit features.
Collaborative filtering
Recommend items other similar users liked. 'Users who liked X also liked Y.'
Matrix factorization. Classic collaborative filtering. User-item interaction matrix decomposed.
Needs interaction data. Cold-start problem for new users and new items.
Scalability. Works at scale with right infrastructure. Spotify, Netflix, YouTube use variants.
Hybrid approaches
Combine content and collaborative. Mitigates cold-start; captures both explicit features and emergent patterns.
Industry standard. Most production recommenders are hybrid.
Ensemble ranking. Multiple models produce scores; meta-model combines.
Neural / AI-era approaches
Two-tower models. User tower and item tower produce embeddings; similarity drives ranking. Scalable for large catalogs.
Transformers. Sequence-aware recommendations. User's interaction history as sequence; predicts next interaction.
Graph neural networks. User-item interactions as graph; GNN propagates preferences. Captures indirect relationships.
LLM-assisted. LLM enrichment of items, generation of natural language explanations, conversational recommendations.
What LLM era specifically added
Better item embeddings. LLM encoders produce richer representations than older embeddings.
Richer semantic matches. 'Like that but more adventurous' — LLM can understand and match on this.
Explanations. Natural language justification for recommendations. 'I recommend X because you liked Y and similar themes of Z.'
Conversational discovery. Chat interfaces replacing pure ranked lists for some use cases. User iterates preferences.
Cold-start for new items. LLM can extract item features from text, images, reviews. Bootstraps new items into recommendation system.
Production architecture
Candidate generation. Fast retrieval of O(hundreds) candidates from O(millions) items. Two-tower or similar.
Ranking. Deep ranking models score candidates considering context, user history, time of day, device. Typically gradient boosting, transformers, or MoE.
Reranking. Business rules, diversity, exploration — applied after main ranking.
Explainability. Why this recommendation? Increasingly important for trust.
Challenges
Cold-start. New users, new items. Content-based and LLM enrichment help.
Feedback loops. Recommendations affect what users see; users click what they see; creates feedback. Exploration mechanisms counter.
Filter bubbles. Personalization can narrow exposure. Diversity requirements balance.
Privacy. Recommendation systems use data; privacy-preserving techniques (federated learning, on-device models) emerging.
Fairness. Recommendations may disadvantage some items or creators. Active research area.
Evaluation
Offline metrics. Precision@K, Recall@K, NDCG, MAP. Predict user engagement.
Online metrics. Click-through rate, engagement, revenue, retention. Ultimately what matters.
A/B testing. Standard for recommendation changes. See A/B testing post.
Counterfactual evaluation. Estimating what would have happened with different recommendations. Increasingly important.
Industry examples
Netflix. Deep personalization; conversational recommendations emerging.
Spotify. Music recommendations; now conversational via AI DJ feature.
Amazon. Product recommendations; LLM-powered shopping assistant emerging.
TikTok. Short video recommendations — arguably most sophisticated in production.
Enterprise uses. Customer recommendation engines for e-commerce; content recommendations for news/media; product recommendations for B2B.